Written by Martin Kardzhilov
Capital is flooding into brain-computer interfaces, humanoid platforms and AI infrastructure at once. The allocation pattern says investors are paying for capability and treating evidence as a downstream problem. That ordering is backwards.
There is no shortage of capital for physical AI. There is a shortage of the specific thing capital needs in order to be deployed responsibly, and the gap between those two statements is where my job lives.
Look at where money actually landed this month. Precision Neuroscience closed an oversubscribed $250 million Series D to expand its clinical programme and build toward commercialisation of an implantable brain-computer interface. The round was co-led by Pershing Square and the Ackman Oxman Institute alongside a life sciences fund — institutional capital rather than strategic venture money, which is a meaningful signal about how the category is being underwritten.
Meanwhile the infrastructure layer is absorbing capital at a rate with no precedent. Data-centre capital expenditure rose 92% year over year in the second quarter against roughly $720 billion of committed 2026 spending across four hyperscalers. Semiconductor equipment sales are forecast at $165.9 billion this year, rising to $201.2 billion in 2027. None of that is discretionary in the ordinary sense; it is the physical precondition for models to exist at all.
Now look at where the humanoid layer sits. Announced orders across the sector exceed $300 million, and announced production capacity at individual manufacturers runs to 12,000 units a year. A thousand-unit deployment spread across a global factory network over seven years is a rounding error against those figures. The capital is real. The deployed fleet is not yet.
This is not a criticism of ambition. It is an observation about how investment theses get tested. The current pattern funds capability first and evidence second, and in robotics those are not sequential problems.
A demonstration can be financed with a slide deck. A fleet cannot. The moment units are deployed at scale, the cost structure stops being about the robot and starts being about everything around it: how often a human has to step in, how long the handoff takes, how reliably the system recovers, and how much supervision the operation needs per shift. Those are the line items that decide whether a deployment is profitable, and they are the line items nobody puts in a pitch.
For a CFO or CEO, this produces an uncomfortable asymmetry. The capital expenditure is knowable to the dollar. The operating variance is not measured at all. I can tell you what a unit costs and I cannot tell you what it earns per hour in a live facility, because the industry has not standardised the measurement.
Three things would change my view of the sector’s maturity. Vendors publishing intervention and recovery rates as routinely as they publish degrees of freedom. Financing structures that price operating evidence rather than prototype risk. And diligence that treats deployment data as a primary asset instead of a post-investment hope.
The money is not the constraint. The measurement is. Until the second catches up with the first, we will keep funding the demonstration and deferring the question of whether it works.
